%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% GENERAL INFORMATION %%%%%%%%%%%%%%%%%%%%%%%%%%%%%
This file contains details on how to prepare the data and estimate the
reduced-form and structural models in Glitz and Vejlin "Learning through
Coworker Referrals" (RED, 2020).
Operating system: CentOS Linux 7 (Core)
Software: Stata/SE 13.1, Matlab, R2017b (9.3.0.713579) 64-bit (glnxa64)
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% ZIP FILE %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
The ZIP-file ProgramsFinal contains 1) a set of programs decscribed in detail
below and 2) a subfolder (\Datafiles), which contains the original microdata
(EstData19802005_AllCounties.dta), the main data set for the reduced form
estimations (mobilitywageturnoverincyespreparationlimit99.dta) and the empirical
moments values and covariance matrix from the data.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% REDUCED FORM ANALYSIS %%%%%%%%%%%%%%%%%%%%%%%%%%%
*************************************** SAS ************************************
To start the data preparation process, run the SAS file "01PrepDataSet" which
creates the Stata data set "EstData19802005_AllCounties.dta" which is the basis
for all empirical work in the paper.
************************************** Stata ***********************************
Run the main master do-file (masterfile.do) which calls the following programs:
- datapreparation.do: Constructs a data set from the original microdata
including some socio-economic and employment history characteristics.
- placebopreparation.do: Constructs data sets that show for each firm in
each year the set of workers who used to work in that firm in the previous and
post x years but not in the year itself.
- matrices.do: Constructs data sets that show for each worker all of
his/her actual and placebo coworkers.
- massmovementindicator.do: Constructs an indicator for movements of
more than one worker from one firm to another between any two subsequent years.
- mobilitypreparationpart1.do: Constructs an intermediate data set used
in the later regression models.
- mobilitypreparationpart2.do: Constructs an intermediate data set used
in the later regression models.
- mobilitysystematic.do: Randomly allocates movers to firms and
calculates exposure measures.
- mobilitydescriptives.do: Provides some first descriptive statistics
(though not for the final estimation sample).
- mobilitywageturnoverpreparation.do: Prepares the data sets used in the
main regressions.
- mobilitywageturnover.do: Runs all the main regression models (for
heterogeneity results, choose the appropriate locals in the command block at the
top of the program).
- mobilitywageturnoverstayers.do: Runs the regression models that
distinguish between the profiles of stayers and the effect due to dynamic
selection.
- mobilitygraphsconvergence.do: Generates the convergence graphs based
on the previous estimation results.
- mobilitydescriptivespost.do: Provides descriptive statistics for the
final estimation sample.
- mobilitywageturnoverVARCOVmatrixlambda.do: Builds on the main
regression file (mobilitywageturnover.do) but extracts the variance-covariance
matrix of the coefficient estimates of our two main models (which are needed in
the structural estimation part).
- mobilitywageturnoverpreparationXX.do: Same as
mobilitywageturnoverpreparation.do only that now the data is prepared for
different maximum network sizes (chosen by setting the local
"networklimitquantile" at the beginning of the file to 50, 75, 90 or 95).
- mobilitywageturnoverXX.do: Runs robustness checks with respect to the
maximum network size allowed (Table C-1).
- mobilitywageturnover_additonal.do: Runs robustness checks with respect
to different market sizes (Table C-2).
Expected computation time: Several days since especially some of the data
preparation files (e.g. matrices.do) take a long time to complete.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%% STRUCTURAL MODEL %%%%%%%%%%%%%%%%%%%%%%%%%%%
Run the main program (main.m). This calls the following programs
- Setup.m: Set up exogeneous values. It sets the random number generator
to Mersenne Twister using a seed of 1.
For estimation we have used both a particle swarm algorithm and a gradient-based
approach. These are called in line 67-68 of main.m. The optimizers calls either
Quadraticfunction.m or QuadraticfunctionParFor.m.
Quadraticfunction.m: Objective function value that calls
- Simulation.m: Simulates worker histories
- Nleq1.m: Solves reservation values- Moments.m: Calculates moments
QuadraticfunctionParFor.m: Objective function value (with parallelizing in the
simultion of worker histories). It calls the same set of programs as
Quadraticfunction.m.
After the estimation/optimal structural parameters are found main.m calls the
following programs.
Momplot.m: Makes plots of model fit
latexTable.m: Used to generate latex output tables
Stderr.m: Calculates std. errors
CounterFac.m: Calculates the counterfactural scenarioes
Expected computation time: This is determined by the number of workers that is
simulated in the estimation. With 5 million workers the expected computation
time for estimation using 70 cores is around 5-10 days. Without estimation
(commenting out lines 67-68) it takes less than 1 hour.